2026-11-08 · America/Los_Angeles · 社区动态 · #16

Moss training has has a very large speed boost in Fizgig

Training is much faster, uses far less memory, and likeness is unchanged. On an RTX 5090, a MOSS-Image LoRA at batch 1 (94 pictures, Muon, 70 epochs) now takes about 20 minutes instead of 49. Other models have had speed boosts too, but Moss the most dramatic. Batch training is now fully supported too. Other changes PyTorch 2.15. With gradient checkpointing on, training uses much less memory, at batch 1 too: Krea 2 22.9 → 14.9 GB, MOSS-Image at batch 3 21.7 → 7.1 GB. Muon is much faster. Its step now handles a LoRA's matrices together instead of one at a time: 12× faster Muon calculation,…

热度 45.3 / 100；排名与评分保留该期记录。